MemeChain: Polysemous Meme-Driven Systems

• MemeChain is a polysemous concept defined across blockchain market dynamics, forensic datasets, and multimodal reasoning frameworks for meme analysis.
• In the Solana context, MemeChain denotes a retail-driven memecoin layer where rapid token issuance and trading reflect high participation amid systemic risks.
• As a dataset and ML framework, MemeChain integrates on-chain and off-chain signals to enable structured forensic analysis and preference-aligned meme generation.

MemeChain is a polysemous research term whose meaning depends on disciplinary context. In recent arXiv literature, it denotes three distinct but related constructs: an emergent layer of the Solana ecosystem organized around rapid, retail-driven memecoin issuance and trading; a large-scale, open-source, cross-chain dataset for meme-coin forensics and risk analysis; and a multimodal Chain-of-Thought architecture for meme understanding and generation. Across these usages, the unifying theme is the treatment of memes or meme coins as structured socio-technical objects whose lifecycle depends on the interaction of on-chain activity, off-chain signaling, multimodal semantics, and collective attention dynamics (Mancino, 4 Dec 2025, Mongardini et al., 28 Jan 2026, Nandi et al., 2024, Li et al., 31 Dec 2025).

1. Terminological scope and conceptual variants

In the Solana market-structure study, MemeChain is defined as “the emergent layer of the Solana ecosystem characterized by rapid, retail-driven token creation and trading activity centered on meme-themed cryptocurrencies.” It is explicitly “rather than a new protocol,” and instead refers to a socio-technical network of memecoin issuances via platforms such as Pump.fun, the on-chain interactions involving those tokens, and the collective market dynamics generated by social hype and speculative flows (Mancino, 4 Dec 2025).

In the forensic-data literature, MemeChain is the name of “the first large-scale, open-source, cross-chain dataset that unifies on-chain token metadata with off-chain web and social signals to support forensic analysis and risk modeling of meme coins.” In that usage, the term refers neither to a market layer nor to a blockchain protocol, but to a research resource spanning Ethereum, BNB Smart Chain, Solana, and Base (Mongardini et al., 28 Jan 2026).

In multimodal machine learning, MemeChain appears as a general Chain-of-Thought framework inspired by SAFE-MEME and as a meme-generation pipeline instantiated by HUMOR. In the former, it is a structured reasoning system that decomposes a meme into sequential sub-questions or hierarchical decisions over image and text. In the latter, it is a hierarchical, multi-path Chain-of-Thought process for generating in-the-wild memes under group-wise human preference alignment (Nandi et al., 2024, Li et al., 31 Dec 2025).

A common misconception is to treat MemeChain as a single protocol or product category. The literature does not support that interpretation. Instead, the term is used across blockchain economics, financial forensics, and multimodal reasoning, with each usage retaining the meme-centric emphasis while changing the object of analysis.

2. MemeChain as a Solana memecoin market layer

Within Solana, MemeChain denotes an overlay produced by memecoin issuance, trading, and social coordination. The defining empirical reference point is Pump.fun during Q4 2024. The study reports token minting share as

$$S_{\text{tokens}}=\frac{\text{Pump.fun\_mints}}{\text{Total\_Solana\_mints}},$$

with an observed peak of approximately

$$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$

or 71.1%. At its height, Pump.fun was therefore responsible for over 70% of all new tokens minted on Solana (Mancino, 4 Dec 2025).

The same study defines DEX transaction share as

$$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$

with an observed range of 0.40–0.674, meaning 40%–67.4% of all on-chain DEX transactions on Solana involved tokens originally created on Pump.fun. Daily active users rose from 60,000 in early October to peaks of 260,000 in November. Under the growth model

$N(t)=N_0\cdot e^{rt},$

with $N_0=60{,}000$, $N(t_1)=260{,}000$, and $t_1\approx 40$ days, the estimated growth rate is approximately $r\approx 0.0366$ day$^{-1}$, or roughly 3–4% daily growth. The paper further defines success rate to major DEXes as

$$p_{\text{success}}=\frac{\text{Tokens\_graduated\_to\_major\_DEX}}{\text{Tokens\_minted\_on\_Pump.fun}},$$

and reports $$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$0, indicating that fewer than 2% of tokens met the volume and liquidity criteria to “graduate” to a major DEX such as Raydium (Mancino, 4 Dec 2025).

These metrics are embedded in a broader account of retail adoption and speculative churn. The study states that barriers to participation collapsed because anyone could launch a token in minutes, and that over 3 million tokens were minted in Q4, averaging seven per minute. It also notes that high transaction counts, constituting 40%–67.4% of all DEX transactions, were paired with comparatively low trading volumes, at less than 50% of DEX volume, which points to many small, rapid trades. The same source associates this with social momentum, hype cycles, volatility, sudden dumps, and a flood of low-quality tokens that may dilute liquidity, fragment order books, and increase slippage across Solana DEXes. Infrastructure stress is also emphasized: market infrastructure had to absorb bursts of traffic amounting to 2–4 million daily transactions on Pump.fun alone (Mancino, 4 Dec 2025).

The paper frames these observations as a dual impact. On one side, lowered barriers, memetic branding, and social coordination catalyzed mass participation. On the other, the concentration of more than 71% of token creation on a single retail platform and a sub-2% graduation rate imply ephemeral value formation and heightened fragility. This suggests that MemeChain, in the Solana sense, is best understood as a measurable market regime rather than merely a cultural trend.

3. MemeChain as a cross-chain forensic dataset

As a dataset, MemeChain was constructed to address the limited observability of meme coins when only single-chain or purely on-chain data are available. The dataset comprises 34,988 meme coins collected across Ethereum, BNB Smart Chain, Solana, and Base. Construction began with 8,852 “verified” meme-coin seeds retrieved via public APIs from CoinMarketCap, CoinGecko, and GeckoTerminal. To capture emergent tokens, the authors scraped DexScreener and CoinSniper, collecting 65,021 unique smart contracts. A two-step classification then combined TF-IDF-derived meme-keyword filtering of token names, using 126 high-value terms, with direct inclusion of pump.fun tokens identified via the “.pump” suffix, producing 28,975 additional meme coins. A three-stage refinement removed stablecoins, large outliers with price greater than \$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$110^7$ USD that were non-meme, and tokens whose names contained “usd,” “wrapped,” or “staked,” yielding 34,988 candidate tokens (Mongardini et al., 28 Jan 2026).

Address validation reduced this set to 31,811 valid meme-coin contracts, or 96.76%. The validated distribution is BSC 15,455, Solana 11,809, Ethereum 3,720, and Base 827. Off-chain integration is central to the dataset design. Token logos were downloaded from the same aggregators, with 15,095 tokens, or 43.14%, including an image. Project websites were normalized, shortened URLs resolved, Linktree handling applied, and non-dedicated domains filtered out, producing 21,113 candidate sites whose raw HTML snapshots were archived when reachable. Social media handles for Telegram, X, and Discord were extracted and validated. WHOIS data were retrieved for 11,819 websites, web accessibility was tested via socket checks and HTTP status codes, and security labels were recorded from Google Safe Browsing and ChainPatrol (Mongardini et al., 28 Jan 2026).

The feature schema spans both on-chain and off-chain modalities. On-chain fields include token_address, chain, token_name, symbol, deploy_date, first and last transaction dates, total transaction count, and market metrics such as price_usd and market_cap where available. Volatility is defined as

$$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$2

Dropout indicators include the “One-Day Meme Coin” flag for inactivity within 24 hours. Off-chain fields include logo_presence, website_url, html_path, domain_metadata, web_access, social_presence, and security_flags (Mongardini et al., 28 Jan 2026).

The descriptive statistics emphasize extreme churn. The abstract identifies 1,801 tokens that cease all trading activity within 24 hours of launch and gives the corresponding share as 5.15%. The detailed metric section defines

$$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$3

Within three months of data collection, 15,099 tokens, or 47.5%, had ceased all activity, and 4,939, or 15.53%, died within 30 days. Web-presence statistics are similarly revealing: 23,755 tokens, or 74.8%, claimed an official website; after cleaning, 21,113 dedicated domains remained; only 10,722, or 50.8%, were reachable by socket; and just 6,766, or 32.1%, returned HTTP 200. The paper also reports that 43.1% of tokens provide a logo and 56.9% lack any visual branding (Mongardini et al., 28 Jan 2026).

The dataset is intended for multimodal scam detection, early survival prediction, cross-chain capital-migration studies, and social-sentiment analysis. Examples given include HTML forensic analysis for reused phishing-kit templates and hidden JavaScript trojans, Siamese-network logo embeddings for brand impersonation, transaction-velocity plus social-media presence plus web accessibility as a real-time risk signal, and comparative fraud prevalence analyses in which BSC exhibits the highest one-day dropout rate at 10.35% versus Solana at 0.09% (Mongardini et al., 28 Jan 2026).

4. MemeChain as a multimodal reasoning framework for meme analysis

In the SAFE-MEME-derived blueprint, MemeChain is a general multimodal Chain-of-Thought framework for meme analysis. The core object is a meme represented as $$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$4, where $$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$5 is the image and $$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$6 is the overlaid text. The framework requires the model to read $$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$7, inspect $$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$8, generate sub-questions about objects, relations, and implied meaning, answer those questions, and then assign a final label $$S_{\text{tokens}} \approx \frac{69{,}046}{\approx 97{,}100}\approx 0.711,$$9 (Nandi et al., 2024).

Two variants are defined. SAFE-MEME-QA generates a list of questions $$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$0, answers each in turn, and then applies either a small rule or a lightweight classifier over $$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$1 to decide the label. SAFE-MEME-H first generates a rich visual description $$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$2 and then performs two-level classification: level 1 determines hateful versus benign, and level 2 refines hateful into explicit versus implicit. Both variants are implemented as multimodal Transformer pipelines with a text encoder, a vision encoder, gated fusion, and a T5-style decoder (Nandi et al., 2024).

The mathematical core includes text encoding

$$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$3

vision encoding

$$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$4

cross-attention using projected queries, keys, and values, and gated fusion

$$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$5

$$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$6

Generation uses a T5-style autoregressive decoder with loss

$$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$7

The hierarchical variant adds protected-group adapters and a two-stage loss, while adapter fine-tuning inserts a two-layer feed-forward adapter between Transformer blocks:

$$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$8

with only $$S_{\text{txs}}=\frac{\text{DEX\_trades\_of\_Pump.fun\_tokens}}{\text{Total\_Solana\_DEX\_trades}},$$9 and layer norms updated. Confidence is defined by negative entropy over final softmax logits (Nandi et al., 2024).

The associated datasets are MHS, with 3,343 memes and splits 2,233/305/805, and MHS-Con, with 1,238 confounders in which the same image is paired with three different texts yielding explicit, implicit, and benign interpretations. Preprocessing includes OCR or human transcription of overlaid text, resizing to $N(t)=N_0\cdot e^{rt},$0, standard normalization, train-time augmentations such as random crop, color jitter, and horizontal flip, SentencePiece tokenization, and ViT image patches. The training defaults are batch size 32, learning rate $N(t)=N_0\cdot e^{rt},$1 for adapters or $N(t)=N_0\cdot e^{rt},$2 for full tuning, weight decay $N(t)=N_0\cdot e^{rt},$3, AdamW, 1,000 warmup steps, roughly 20K total steps, and early stopping on validation macro-F1 (Nandi et al., 2024).

The reported benchmark values are specific. SAFE-MEME-QA yields approximately 0.54 macro-F1 on MHS versus 0.49 for CLIP (FT), and approximately 0.62 macro-F1 on MHS-Con versus 0.58 for T5 (FT). SAFE-MEME-H peaks at approximately 0.55 macro-F1 on MHS. The worked example involving a dishwasher rack filled with women’s shoes and the text “She does what she’s made for” illustrates how the QA path can infer implicit mockery and how the hierarchical path can assign hateful then implicit with probabilities 0.72 and 0.65, respectively. The implementation notes identify class imbalance, hallucination in QA, and the fuzzy explicit-versus-implicit boundary as recurring failure modes, and recommend balanced sampling, group-specific adapter pre-training, constrained question templates or retrieval of external definitions, hierarchical crowdsourcing guidelines, soft labels, and entropy-based confidence filtering (Nandi et al., 2024).

5. MemeChain as a pipeline for in-the-wild meme generation

In the HUMOR framework, MemeChain refers to a meme-generation system that treats humor production as hierarchical, multi-path reasoning followed by group-wise preference alignment. The process begins with template-level intent inference: given an input image $N(t)=N_0\cdot e^{rt},$4 and an optional user context $N(t)=N_0\cdot e^{rt},$5, the model generates $N(t)=N_0\cdot e^{rt},$6 template-level intents $N(t)=N_0\cdot e^{rt},$7. Conditioned on each intent, it explores multiple context-level paths $N(t)=N_0\cdot e^{rt},$8. During training, a gold punchline $N(t)=N_0\cdot e^{rt},$9 is available, and one reasoning trace is anchored by selecting

$N_0=60{,}000$0

where $N_0=60{,}000$1 is the implied reasoning extracted from the gold caption (Li et al., 31 Dec 2025).

The framework formalizes humor quality through a group-relative score $N_0=60{,}000$2. Under the condition that a set of good paths $N_0=60{,}000$3 has total mass at least $N_0=60{,}000$4 and the worst gap between a bad path and the best path is at most $N_0=60{,}000$5, Proposition 1 gives the lower bound

$N_0=60{,}000$6

This is used to justify the claim that multi-path exploration can preserve diversity while anchoring retains a non-negligible probability mass on at least one high-quality path (Li et al., 31 Dec 2025).

Preference alignment is group-wise and pairwise. Meme space is partitioned into template groups $N_0=60{,}000$7, where captions share the same base image. The observed pairwise label $N_0=60{,}000$8 is modeled as

$N_0=60{,}000$9

and the learned scorer $N(t_1)=260{,}000$0 induces

$N(t_1)=260{,}000$1

Training minimizes binary cross-entropy over annotated pairs. Proposition 2 states rank consistency, and Proposition 3 states noise robustness when annotation error rate is $N(t_1)=260{,}000$2, with flips concentrating among small-margin pairs (Li et al., 31 Dec 2025).

Reinforcement learning is implemented as group-wise relative policy optimization. For each group $N(t_1)=260{,}000$3, an EBC-derived target distribution $N(t_1)=260{,}000$4 is defined, and the objective is

$N(t_1)=260{,}000$5

Proposition 4 provides a monotonic-improvement-style guarantee: if the expected KL divergence from the reference policy is bounded by $N(t_1)=260{,}000$6, then expected humor cannot degrade by more than a term proportional to $N(t_1)=260{,}000$7 (Li et al., 31 Dec 2025).

The reported experimental configuration uses a training corpus of approximately 3.7K real in-the-wild memes annotated with OCR text positions, emotion and intention tags, and pairwise funniness labels across five difficulty tiers. Evaluation includes human judgments on Humor, Readability, Relevance, and Originality using a 1–5 scale; Context-Swap Distance; Reference Similarity via bge-base-en-v1.5; Human-Rate as the percentage of outputs labeled human-made by Gemini-2.5-pro; and group-wise ranking via Gemini-2.5-pro and human MaxDiff. On the excerpted comparison against Qwen2.5-7B, HUMOR-CoT raises Humor from 2.39 to 2.68, Readability from 3.35 to 3.70, Relevance from 2.91 to 3.50, Originality from 2.57 to 2.90, Context-Swap Distance from 0.564 to 0.590, Reference Similarity from 0.576 to 0.640, and Human-Rate from 75.7% to 91.5%. The reward model based on Keye-VL-8B achieves Kendall $N(t_1)=260{,}000$8 across five unseen templates, and a HUMOR-RL preview raises Humor to 2.83 and Human-Rate to 92.3% (Li et al., 31 Dec 2025).

The paper also contrasts HUMOR-CoT with alternative reasoning strategies. Single-path CoT is described as “safe but dull” and scores Humor 1.87; Self-Improve scores Humor 2.38; Subquestion scores Humor 1.85; and hierarchical multi-path plus anchoring is presented as the best balance of diversity and quality. The stated extension domain includes stylistic image captioning, aesthetic image retouching, personalized social-media posts, visual poetry, ad-copy generation, and emoji synthesis (Li et al., 31 Dec 2025).

6. Cross-cutting themes, misconceptions, and research directions

Across its different meanings, MemeChain consistently links high-velocity production with weak observability and noisy signals. In the Solana market-layer usage, the central problem is systemic risk under retail-driven speculative concentration. In the dataset usage, the central problem is that off-chain legitimacy signals such as websites, logos, and social accounts are indispensable but often missing, abandoned, or adversarial. In the multimodal-ML usages, the central problem is that meme interpretation and generation require structured intermediate reasoning rather than direct input-output mapping (Mancino, 4 Dec 2025, Mongardini et al., 28 Jan 2026, Nandi et al., 2024, Li et al., 31 Dec 2025).

A second recurring theme is that high activity does not imply durable value. The Solana study states explicitly that dominance in on-chain metrics does not equate to sustainable value, despite rapid user growth and market share. The dataset paper shows this at cross-chain scale through one-day dropouts, 30-day deaths, three-month inactivity, and weak web persistence. The analytical implication is not that all meme-centric systems are fraudulent or transient, but that churn, abandonment, and speculative feedback loops are measurable first-order properties of the domain (Mancino, 4 Dec 2025, Mongardini et al., 28 Jan 2026).

The literature also outlines concrete future directions. For Solana-style MemeChain monitoring, proposed measures include on-chain health indices combining $N(t_1)=260{,}000$9, $t_1\approx 40$0, $t_1\approx 40$1, and token concentration measures; dynamic graduation thresholds or reputation mechanisms; layer-2 risk controls or insurance pools; and monitoring user growth rates against transaction-to-volume ratios. For the dataset, suggested extensions include continuous monitoring of web content, integration of holder-distribution and wallet-clustering features, expansion to additional chains such as Avalanche and Polygon, and inclusion of real-time price feeds. For multimodal reasoning, recommended improvements include external-definition retrieval, constrained question templates, group-specific adapter pre-training, hierarchical annotation guidelines, and confidence calibration through entropy-based filtering. For generation, the group-wise framework is explicitly proposed as a general paradigm for open-ended, human-aligned multimodal generation beyond memes (Mancino, 4 Dec 2025, Mongardini et al., 28 Jan 2026, Nandi et al., 2024, Li et al., 31 Dec 2025).

Taken together, these strands indicate that MemeChain is best understood as a family of formalizations for meme-driven systems rather than a single object. Depending on context, it names a speculative market regime, a multimodal forensic corpus, or a structured reasoning-and-generation architecture. The shared research objective is to make meme-centered phenomena legible under technical scrutiny: economically through on-chain metrics, forensically through cross-modal evidence, and computationally through explicit reasoning traces and preference-aligned generation.